XiaoMi / mobile-ai-bench

Benchmarking Neural Network Inference on Mobile Devices
Apache License 2.0
353 stars 57 forks source link

使用Precision 结果不是1就是0 #22

Closed authorfu closed 5 years ago

authorfu commented 5 years ago

使用imageNet的图片及其它测试,置信度结果不是0.0就是1.0 期望比如有0.98这种的置信度

git pull bash tools/benchmark.sh --benchmark_option=Precision --input_dir=/home/fu/mace/mobile-ai-bench/input_less --executors=MACE --device_types=CPU --model_names=MobileNetV1 --target_abis=armeabi-v7a

测试方法:input_dir只放一张图片,运行上述脚本, cat output/precision_report.csv

ILSVRC2012_val_00000005.JPEG 0.0 ILSVRC2012_val_00000007.JPEG 1.0 ILSVRC2012_val_00000009.JPEG 1.0

测试Preformance,看上去正常 android-ndk-r15c

QQ号 744734320

lee-bin commented 5 years ago

如之前沟通,若需要输出单张图片的置信度可以修改此文件 https://github.com/XiaoMi/mobile-ai-bench/blob/master/aibench/benchmark/imagenet/imagenet_postprocessor.cc